1. Field of the Invention
The present invention relates to a parallel computing system which is constituted by a group of information processing apparatuses each having at least one processor, a memory, and a cache memory as main components, and a connection route for connecting the group of information processing apparatuses.
2. Related Background Art
When a group of information processing apparatuses (to be referred to as nodes hereinafter) each having at least one processor and a memory are connected for the purpose of exchanging data among the nodes, a method of connecting them using various kinds of LANs, and a method of connecting them on the memory address level without using a LAN, are available. As an example of the latter method, the assignee of the present invention proposed an information processing apparatus using an optical wavelength multiplexing method in U.S. Pat. No. 5,602,663, now U.S. Pat. No. 5,602,663. According to this method, by using a plurality of identifiable wavelengths, different data transfer operations can be realized at the same time among a plurality of nodes.
Furthermore, that patent discloses an information apparatus in which each node sends additional information associated with data transfer to an arbiter together with a connection route request prior to the data transfer, and the arbiter sends the received information to a connection end node via an arbitration signal line. The node which received the connection request simultaneously realizes set-up control of the connection route and preparation control of data to be transmitted/received, thus improving the data transfer efficiency after the connection route is set.
When data are shared on the memory address level like in the above-mentioned system, each node normally comprises a cache memory to decrease the number of times of access to the data transfer path so as to improve the use efficiency of data in each node. In this case, since only one data transfer operation is allowed at a time on the data transfer path, the cache memories are snooped to maintain the consistency of data in them.
However, in the system using the optical wavelength multiplexing method, since different data transfer operations can be simultaneously realized among a plurality of nodes using a plurality of wavelengths, it is impossible to simultaneously snoop the plurality of data transfer operations from the nodes. For this reason, it is difficult to use, in each node, a cache memory for holding data in memories in other nodes.
For this reason, data requests for data on memories in other nodes constantly appear on an inter-node connection route, resulting in low computing efficiency.